A multi-center applicable brain structure automatic segmentation method for human brain magnetic resonance images

Abstract

The application relates to a multi-center applicable human brain magnetic resonance image brain structure automatic segmentation method, and the segmentation method specifically comprises the following steps: acquiring MRI data of different platforms, and performing N4 bias field correction and image standardization processing on the MRI data; taking U-Net as a basic framework, fusing a self-attention mechanism module, and constructing an MRI brain image structure segmentation network model suitable for multiple platforms; inputting the processed MRI data into the segmentation network model, training the segmentation network model by using transfer learning and domain adaptation technology, estimating the uncertainty of the output result, and further training the model by using high-quality unlabeled data screened out, so that the human brain magnetic resonance image brain structure automatic segmentation is realized. Compared with the prior art, the application can accurately and quickly segment the brain structure of different magnetic resonance imaging platforms, provides a more universal, efficient and reliable tool for the field of brain image analysis, and helps more extensive application and research exploration.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and medical imaging technology, and in particular to an automatic brain structure segmentation method for human brain magnetic resonance images applicable to multiple centers. Background Technology

[0002] Brain science research plays a crucial role in exploring human evolution, development, and the treatment of neurological diseases. Magnetic resonance imaging (MRI), as an ideal non-invasive imaging technique, can provide high-resolution and multimodal imaging of the brain. Segmentation of brain regions from MRI data is a key step in analyzing brain function and metabolic processes.

[0003] Fine-grained brain segmentation is a key technology in neuroscience and medicine, aiming to divide brain imaging data into different anatomical or functional regions. This task is of paramount importance in advancing scientific research and clinical applications. In humans, fine-grained brain segmentation has direct applications in clinical diagnosis and treatment, providing crucial lesion information in diseases such as Alzheimer's disease, Parkinson's disease, and brain tumors. These segmentation techniques not only contribute to a better understanding of brain structure and function but also provide important evidence for developing personalized treatment plans.

[0004] By focusing on the study of human brain MRI data, we can gain in-depth insights into brain function and pathological changes, providing more accurate and reliable support for improving clinical diagnosis and treatment methods. This not only advances basic research in brain science but also brings new hope for the treatment of neurological diseases.

[0005] Fine-grained brain region segmentation faces numerous challenges, primarily stemming from the high complexity of brain structure and individual variability. The brain contains numerous intricate anatomical regions, whose boundaries are often blurred and difficult to distinguish. Individual variability in brain structure further complicates segmentation. Furthermore, acquiring high-quality brain imaging data requires expensive equipment and technology, while manually annotating this data is time-consuming and susceptible to subjective bias. Therefore, developing algorithms capable of automatically and accurately segmenting brain regions has become a crucial task in this field.

[0006] In performing fine brain region segmentation on single-modality T1-weighted MRI images, the heterogeneity of data sources presents significant challenges. Differences in imaging parameters due to different MRI equipment and scanning protocols, inconsistent image quality, different standardization methods, and variations in data formats and preprocessing procedures can all lead to inconsistent representations of the same anatomical region across different images, making it difficult for segmentation algorithms to maintain stable performance. Furthermore, different research teams may use different annotation standards and methods, further increasing the inconsistency of training data and affecting the generalization ability of segmentation models.

[0007] Currently, various brain structure segmentation tools designed for specific MRI modalities exist, and the development of deep learning algorithms has also contributed to improving the accuracy of human brain structure segmentation. However, the performance of these methods often degrades significantly when processing data from different platforms or scanning protocols. Accurately segmenting fine brain regions from images acquired from multiple MRI platforms using a single tool remains a challenge.

[0008] Furthermore, existing segmentation algorithms typically only perform segmentation and lack self-checking and correction capabilities, resulting in inconsistent accuracy and reliability of the segmentation results. Because these algorithms cannot automatically identify and correct segmentation errors, significant human intervention is required for subsequent checks and corrections. This reliance on manual review is not only extremely time-consuming but also highly inefficient when processing large-scale data. Each segmentation result requires careful expert review to ensure it meets expected accuracy and precision. This process is not only tedious but also susceptible to subjective human factors, leading to inconsistent segmentation quality. Moreover, manual checking and correction increase workload, limiting the widespread use of these algorithms in practical applications. In clinical and research settings, this manual checking and correction process not only wastes valuable human resources but also prolongs data processing cycles, potentially impacting diagnostic and research progress. The limitations of this approach are particularly pronounced when rapid processing and analysis of large amounts of data are required.

[0009] In general, existing segmentation technologies face the following challenges: 1. Lack of flexibility and universality. Existing tools are designed for specific platforms and are not suitable for cross-platform applications. 2. Lack of automated quality assessment mechanisms. This means that in cases of incorrect segmentation, tedious manual checks and corrections are still required.

[0010] In summary, medical brain imaging structural segmentation technology faces challenges in terms of model flexibility and universality across different platforms, as well as automated quality assessment and adjustment. Therefore, how to introduce more flexible transfer learning methods and domain adaptation techniques into the model to achieve more accurate and robust brain structural segmentation in complex cross-domain data environments has become a crucial technical problem that urgently needs to be solved. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of the existing technology by providing an automatic brain structure segmentation method for human brain magnetic resonance images applicable to multiple centers. By designing a robust deep learning model, segmentation can be performed directly on the raw image data, eliminating the need for complex preprocessing, simplifying the usage process, lowering the technical threshold, and maintaining high-precision segmentation performance.

[0012] The objective of this invention can be achieved through the following technical solutions:

[0013] This invention provides an automatic brain structure segmentation method for human brain magnetic resonance images applicable to multiple centers, characterized by comprising the following steps:

[0014] S1: Acquire MRI data from different platforms and perform N4 bias field correction and image normalization on the MRI data;

[0015] S2: Based on U-Net, and incorporating a self-attention mechanism module, a network model for MRI brain image structure segmentation adapted to multiple platforms is constructed.

[0016] S3: Input the MRI data processed in S1 into the MRI brain image structure segmentation network model constructed in S2, and train the MRI brain image structure segmentation network model using transfer learning and domain adaptation techniques.

[0017] S4: Uncertainty estimation is performed on the output structure of the MRI brain image structure segmentation network model in S3, and the segmentation results are evaluated and screened. The high-quality unlabeled data selected are used to further train the model, and the trained model is used to achieve automatic segmentation of brain structures in human brain MRI images.

[0018] Further, in S1, the standardization process converts the original image V into a standardized MRI image V. norm The calculation formula is as follows:

[0019]

[0020] in, and These represent the 1st and 99th percentiles of the intensity values ​​in the input image, i.e., the lower and upper limits of the intensity values; V is the original image, V norm This is the image after deviation normalization.

[0021] Furthermore, in S2, the U-Net includes five sets of symmetrical encoding blocks and decoding blocks. Each encoding block consists of two 3x3x3 convolutions, followed by a batch normalization layer, an exponential linear unit activation function, and a 2x2x2 max pooling operation, with a stride of 2.

[0022] The first coding block outputs 32 feature channels, and the number of feature channels doubles with each subsequent coding block.

[0023] Furthermore, each decoded block consists of two 3x3x3 convolutions, followed by a batch normalization layer, an exponential linear unit activation function, and a transposed convolutional layer;

[0024] In the final decoding stage, a 1x1x1 convolutional layer and a sigmoid activation function are used to map the 32 feature channels to a 96-ary probability map.

[0025] Furthermore, in S2, the self-attention mechanism module simulates long-distance and global contextual dependencies in feature representation by adding a non-local attention layer at the center of the network model, and calculates a weighted saliency map at each location on the feature map in order to more effectively distinguish different brain regions with different contrast characteristics and distributions.

[0026] Furthermore, in S3, the Dice loss function and adaptive batch normalization are used to train the MRI brain image structure segmentation network model, with 50 training epochs, a batch size of 16, and a learning rate of 10 in both the source and target domains. -4 and 10 -5 ;

[0027] An adaptive batch normalization strategy is adopted to dynamically update the statistical parameters of the batch normalization layer during the inference phase, thereby improving the model's cross-platform generalization performance.

[0028] Furthermore, in S4, approximate Bayesian inference is used to estimate the uncertainty of the MRI brain image structure segmentation network model. The specific process is as follows: a dropout layer is introduced to obtain a measure of the uncertainty of the model when generating segments. The dropout layer is kept active during the inference phase. N independent samples are collected through Monte Carlo sampling, and the pixel-level uncertainty of each sample is calculated using the following formula:

[0029]

[0030] Where, p t This represents the output probability of the sigmoid function in the final layer of the model during the inference phase. is the average predicted value of all inference results; T is the number of Monte Carlo samplings, and T = 10 is set.

[0031] Furthermore, in S4, based on uncertainty estimation, a semi-supervised process is used to screen high-quality segmentation results and filter out prediction results with high uncertainty and relatively unreliable results to achieve quality assessment.

[0032] Furthermore, the semi-supervised process specifically involves: evaluating the model on all unlabeled data, then selecting the predictions with the lowest uncertainty based on Monte Carlo quality assessment, and adding them as pseudo-labels to the training set in the next iteration.

[0033] Furthermore, the semi-supervised process employs a hybrid loss function, the specific formula of which is as follows:

[0034]

[0035]

[0036] Where, x i For the current small batch containing the true labeled y i The total number of labeled data points is N; x j For the model in the previous iteration, the parameters θ t-1 The predicted pseudo-labeled data points, whose uncertainty is calculated to be below the threshold ζ by the MCQA module, have a total number of M; f(x j ;θ t ) represents the model's prediction of the input x with parameters θ; λ Dice The hyperparameter α(t) controls the weights of the Dice loss term; α(t) is used to balance the relative contributions of the true and false labels in the algorithm; T sl The preset threshold number of iterations controls when to start introducing pseudo-labels; λ sl The balancing coefficient determines the weight of the pseudo-label loss in the total loss; t represents the current iteration number.

[0037] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0038] 1. This invention is based on deep learning technology and trains the model using a large amount of data, enabling it to capture and learn complex features in image data. Compared with traditional rule-based methods and toolkits, deep learning methods have higher processing speed and stronger generalization ability. Especially when dealing with large-scale datasets, deep learning algorithms can significantly improve segmentation efficiency and quickly output high-quality segmentation results.

[0039] 2. This invention eliminates the need for complex preprocessing. Traditional MRI image segmentation typically requires complex preprocessing steps, such as denoising and registration, to improve image quality and consistency. These preprocessing steps are not only time-consuming but may also introduce additional errors. This invention, through the design of a robust deep learning model, can perform segmentation directly on the raw image data, eliminating the need for complex preprocessing, simplifying the usage process, lowering the technical threshold, and maintaining high-precision segmentation performance.

[0040] 3. This invention possesses a certain ability to repair noise in pseudo-labels. In semi-supervised learning, the model needs to be trained using a large amount of unlabeled data and a small amount of labeled data. The unlabeled data is used to generate pseudo-labels to assist training. However, pseudo-labels may contain noise, which can affect the model's training performance. The network model in this invention selects the pseudo-label with the highest confidence level to add to the training data each time, thus correcting the noise in the pseudo-labels and ensuring the stability of the training process and the accuracy of the segmentation results, thereby improving the model's robustness.

[0041] 4. This invention acquires prior knowledge through transfer learning, enabling the model to achieve robust segmentation across multiple domains and platforms. Transfer learning allows the model to learn useful features and knowledge from image data of one domain or platform and apply this knowledge to another domain or platform. This allows the invention to achieve consistent segmentation results on different MRI platforms. Through transfer learning, the model can quickly adapt to new data sources, significantly improving cross-platform segmentation robustness, reducing the impact of differences in equipment and imaging parameters, and ensuring the consistency and reliability of segmentation results across different platforms. Attached Figure Description

[0042] Figure 1 A structural framework diagram of an automatic brain structure segmentation method for human brain magnetic resonance images applicable to multiple centers;

[0043] Figure 2 A schematic diagram of a multi-center applicable automatic brain structure segmentation method for human brain magnetic resonance imaging;

[0044] Figure 3 This is a schematic diagram of the segmentation results and 3D rendering of 96 brain structures in Example 1;

[0045] Figure 4 This is a schematic diagram of the segmentation results of 96 brain structures in a multi-center study in Example 1. Detailed Implementation

[0046] The following examples illustrate specific implementations of the present invention. These examples are carried out based on the solution described in the present invention, and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following examples.

[0047] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Any structural / module names, control modes, algorithms, processes, or composition ratios not explicitly stated in this technical solution are considered common technical features disclosed in the prior art.

[0048] Example 1

[0049] This embodiment provides an automatic brain structure segmentation method for human brain magnetic resonance images applicable to multiple centers, such as... Figure 1 As shown, it includes the following steps:

[0050] S1: To reduce the differences between different scanning platforms, a consistent preprocessing procedure is performed on all datasets. Specifically, MRI data obtained from multiple MRI platforms, including Canon, GE, Lowin, Philips, Siemens, and United Imaging, are collected. The N4 bias field correction algorithm is used to perform bias field correction on the MRI data, and then each image is standardized.

[0051] The standardization process converts the original image V into a standardized MRI image V. norm The calculation formula is as follows:

[0052]

[0053] in, and These represent the 1st and 99th percentiles of the intensity values ​​in the input image, i.e., the lower and upper limits of the intensity values; V is the original image, V norm This is the image after deviation normalization.

[0054] When processing MRI data, the voxel size was adjusted to 1 mm to ensure the stability of the results. Furthermore, data augmentation strategies were employed to enable the network to learn invariant features from data presenting variations across different subjects and protocols. Transformations were performed on each input, including random rotation (±10 degrees), random scaling (90%–110%), and random translation (maximum ±10 pixels in each axis). For example, rotation and translation operations can simulate the representation of subjects in different coordinate systems and different head orientations. Scaling operations can simulate inconsistencies in brain volume size that occur at different sites and even within the same subject during brain development.

[0055] S2: Using U-Net as the basic framework and integrating the self-attention mechanism module, a multi-platform MRI brain image structure segmentation network model (BREN) is constructed.

[0056] like Figure 2As shown, the U-Net comprises five symmetrical encoding and decoding blocks, responsible for feature compression and expansion during the data feedforward process, respectively. Each encoding block consists of two 3x3x3 convolutions, followed by a batch normalization layer, an Exponential Linear Unit (ELU) activation function, and a 2x2x2 max-pooling operation with a stride of 2. The first encoding block outputs 32 feature channels, doubling the number of channels with each subsequent encoding block. Each decoding block consists of two 3x3x3 convolutions, followed by a batch normalization layer, an ELU activation function, and a transposed convolutional layer. In the final decoding stage, a 1x1x1 convolutional layer and a sigmoid activation function map the 32 feature channels to a 96-ary probability map for soft segmentation of different brain regions for each pixel. The decoding blocks are structurally symmetrical to their corresponding encoding blocks.

[0057] By integrating a self-attention mechanism module, BREN's cross-domain task performance was significantly enhanced. This mechanism module simulates long-range and global contextual dependencies in feature representations by adding a non-local attention layer at the center of the network, and computes a weighted saliency map at each location on the feature map to more effectively distinguish different brain regions with different contrast characteristics and distributions.

[0058] S3: Input the processed MRI images from S1 into the MRI brain image structure segmentation network model in S2, and train the MRI brain image structure segmentation network model using transfer learning and domain adaptation techniques.

[0059] The BREN network model was trained using the Dice loss function and an adaptive matrix estimation optimizer, with 50 training epochs, a batch size of 16, and learning rates of 10⁻¹⁰ in both the source and target domains. 4 and 10 -5 .

[0060] To overcome the domain shift problem encountered by deep learning models in cross-domain applications, an Adaptive Batch Normalization (AdaBN) strategy is used to dynamically update the statistical parameters of the batch normalization (BN) layer during the inference phase, thereby improving the model's cross-platform generalization performance. AdaBN dynamically updates the statistical parameters in the BN layer for a specific target domain during the inference phase, thus enhancing the model's generalization performance. In brain region segmentation tasks, AdaBN enables the model to adapt to differences in MRI signal intensity across different platforms without altering its architecture or hyperparameters.

[0061] S4: Uncertainty estimation is performed on the output structure of the MRI brain image structure segmentation network model in S3, and the segmentation results are evaluated and screened. The high-quality unlabeled data selected are used to further train the model, and the trained model is used to achieve automatic segmentation of brain structures in human brain MRI images.

[0062] To understand the internal evaluator variability exhibited in BREN and to provide clinicians with valuable insights into model interpretability, a dropout layer was introduced into the BREN network model to obtain a measure of the model's uncertainty in generating segments.

[0063] Uncertainty in predictions within deep neural networks can be broadly categorized into two types: incidental uncertainty and cognitive uncertainty. Incidental uncertainty reflects inherent noise in the input test data, and increasing the amount of training data is generally insufficient to reduce this type of uncertainty. Conversely, cognitive uncertainty involves the uncertainty of model parameters, reflecting a lack of knowledge about the optimal model. In deep learning networks, cognitive uncertainty typically arises from a lack of training data in certain input domain regions and decreases as the diversity of training data distribution increases.

[0064] An approximate Bayesian inference method was used to estimate the uncertainty of an MRI brain image structure segmentation network model. The specific process was as follows: a dropout layer was introduced to obtain a measure of the model's uncertainty during segmentation generation; the dropout layer was kept active during the inference phase; N independent samples were collected using Monte Carlo sampling; and the pixel-level uncertainty for each sample was calculated using the following formula:

[0065]

[0066] Where pt is the output probability of the sigmoid function of the final layer of the model during the inference phase; is the average predicted value of all inference results; T is the number of Monte Carlo samplings, and T = 10 is set.

[0067] Based on uncertainty estimation, a semi-supervised process is used to screen high-quality segmentation results and filter out predictions with high uncertainty and relatively unreliable results to achieve quality assessment. Specifically, the semi-supervised process involves: evaluating the model on all unlabeled data, then using Monte Carlo Quality Assessment (MCQA) to select predictions with the lowest uncertainty, which are then used as pseudo-labels and added to the training set in the next iteration.

[0068] The semi-supervised process uses a hybrid loss function, the specific formula of which is as follows:

[0069]

[0070]

[0071] Where, x i For the current small batch containing the true labeled y i The total number of labeled data points is N; x i For the model in the previous iteration, the parameters θ t-1 The predicted pseudo-labeled data points, whose uncertainty is calculated to be below the threshold ζ by the MCQA module, have a total number of M. f(x) j ;θ t ) represents the model's prediction of the input x with parameters θ. λ Dice This is a hyperparameter that controls the weights of the Dice loss term. α(t) is used to balance the relative contributions of the true and false labels in the algorithm.

[0072] The specific process of the relative contribution of the true label and the false label in the α(t) balancing algorithm is as follows: α(t) in the first stage 0 <t<T s1 In this method, the network is trained directly using labeled data in the target domain;

[0073] In the second stage T sl ≤t≤T epoch In this process, unlabeled data with pseudo-labels participates additionally in the semi-supervised learning process through the determination of the balancing coefficient α(t) to contribute to the loss function;

[0074] Among them, T epoch This represents the total number of training periods, and t represents the current period. Let T be the number of training periods. sl =20, λ Dice =1,λ sl =0.5, and ζ is the minimum uncertainty value in each mini-batch.

[0075] like Figure 3 and Figure 4 As shown, the segmentation results using the network model in this embodiment are presented. The results demonstrate that the BREN network model can stably segment MRI data from different platforms and instruments, exhibiting excellent segmentation performance on all platforms. This advantage gives BREN significant superiority in cross-platform applications, particularly on various MRI platforms such as Canon, GE, Lowin, Philips, Siemens, and United Imaging, where it achieves efficient and accurate brain region segmentation. Furthermore, traditional toolkits (such as registration-based segmentation methods) require significant time for preprocessing and manual checks, while BREN, relying on its powerful deep learning capabilities and automatic quality assessment module, can quickly complete segmentation tasks and ensure the consistency and reliability of results, providing strong support for neuroimaging research and clinical applications.

[0076] The above description of the embodiments is provided to enable those skilled in the art to understand and use the invention. It will be apparent to those skilled in the art that various modifications can be made to these embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the invention should be within the protection scope of the present invention.

Claims

1. A method for automatic brain structure segmentation from human brain magnetic resonance imaging (MRI) images applicable to multiple centers, characterized in that, Includes the following steps: S1: Acquire MRI data from different platforms and perform N4 bias field correction and image normalization on the MRI data; S2: Based on U-Net and incorporating a self-attention mechanism module, this model uses transfer learning and domain adaptation techniques to construct an MRI brain image structure segmentation network model that is adaptable to multiple platforms. S3: Input the MRI data processed in S1 into the MRI brain image structure segmentation network model constructed in S2, and train the MRI brain image structure segmentation network model using transfer learning and domain adaptation techniques. S4: Uncertainty estimation is performed on the output structure of the MRI brain image structure segmentation network model in S3, and the segmentation results are evaluated and screened. The high-quality unlabeled data selected are used to further train the model, and the trained model is used to achieve automatic segmentation of brain structures in human brain MRI images. In S4, approximate Bayesian inference is used to estimate the uncertainty of the MRI brain image structure segmentation network model. The specific process is as follows: a dropout layer is introduced to obtain a measure of the uncertainty of the model when generating segments. The dropout layer is kept active during the inference phase. N independent samples are collected through Monte Carlo sampling, and the pixel-level uncertainty of each sample is calculated using the following formula: in, This represents the output probability of the sigmoid function in the final layer of the model during the inference phase. The average predicted value of all inference results; T is the number of Monte Carlo samplings, and T=10 is set. In S4, based on uncertainty estimation, a semi-supervised process is used to screen high-quality segmentation results and filter out prediction results with high uncertainty and relatively unreliable results in order to achieve quality assessment. The semi-supervised process specifically involves: evaluating the model on all unlabeled data, then selecting the predictions with the lowest uncertainty based on the Monte Carlo quality assessment, and using them as pseudo-labels to add to the training set in the next iteration; The semi-supervised process employs a hybrid loss function, the specific formula of which is as follows: in, For the current small batch containing real annotations The total number of labeled data points is N ; For the model parameters in the previous iteration The predicted pseudo-labeled data points are then analyzed, and the uncertainty is calculated to be below a threshold using the MCQA module. The total number is ; This represents the model's prediction of the input x with respect to the parameter θ; The hyperparameters used to control the weights of the Dice loss terms; Used to balance the relative contributions of real labels and pseudo labels in the algorithm; The preset threshold number of iterations controls when to start introducing pseudo-labels; The balancing coefficient determines the weight of the pseudo-label loss in the total loss; t This indicates the current iteration number.

2. The method for automatic brain structure segmentation in multi-center applicable human brain magnetic resonance images according to claim 1, characterized in that, In S1, the standardization process will transform the original image Converted to standardized MRI images The calculation formula is as follows: in, and These represent the 1st and 99th percentiles of the intensity values ​​in the input image, i.e., the lower and upper limits of the intensity values; For the original image, This is the image after deviation normalization.

3. The method for automatic brain structure segmentation in multi-center applicable human brain magnetic resonance images according to claim 1, characterized in that, In S2, the U-Net includes five sets of symmetrical coding blocks and decoding blocks. Each coding block consists of two 3x3x3 convolutions, followed by a batch normalization layer, an exponential linear unit activation function, and a 2x2x2 max pooling operation, with a stride of 2. The first coding block outputs 32 feature channels, and the number of feature channels doubles with each subsequent coding block.

4. The method for automatic brain structure segmentation in multi-center applicable human brain magnetic resonance images according to claim 3, characterized in that, Each decoded block consists of two 3x3x3 convolutions, followed by a batch normalization layer, an exponential linear unit activation function, and a transposed convolutional layer. In the final decoding stage, a 1x1x1 convolutional layer and a sigmoid activation function are used to map the 32 feature channels to a 96-ary probability map.

5. The method for automatic brain structure segmentation in multi-center applicable human brain magnetic resonance images according to claim 1, characterized in that, In S2, the self-attention mechanism module simulates long-distance and global contextual dependencies in feature representation by adding a non-local attention layer at the center of the network model, and calculates a weighted saliency map at each location on the feature map in order to more effectively distinguish different brain regions with different contrast characteristics and distributions.

6. The method for automatic brain structure segmentation in multi-center applicable human brain magnetic resonance images according to claim 1, characterized in that, In S3, the Dice loss function and adaptive batch normalization were used to train the MRI brain image structure segmentation network model, with 50 training epochs, a batch size of 16, and a learning rate of 10 in both the source and target domains. -4 and 10 -5 ; An adaptive batch normalization strategy is adopted to dynamically update the statistical parameters of the batch normalization layer during the inference phase, thereby improving the model's cross-platform generalization performance.

Images